Investigating Words

You can investigate relations of words with other words or with topics in topicwizard.

The fundamental trick is that you can interpret topic-word importances/probabilities as word embeddings, and this is in fact what people do who train LSI word embeddings.

For example:

# SVD is just the underlying mathematical algorithm for Latent Semantic Indexing/Allocation
from sklearn.decomposition import TruncatedSVD
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.pipeline import make_pipeline

vectorizer = CountVectorizer()
# Word embeddings with 100 dimensions
svd = TruncatedSVD(n_components=100)
pipeline = make_pipeline(vectorizer, svd)
pipeline.fit(corpus)

# Your word embeddings are the components of the topic model transformed
# You can virtually use any topic model like this.
embeddings = svd.components_.T
# THis is the order of words in the embedding matrix
vocab = vectorizer.get_feature_names_out()

This means that you can also use topicwizard for investigating certain word embedding models.

Web App

When you click ‘Words’ on the navigation bar of the web application you will be presented with this screen.

This is mainly dedicated to interpreting word associations and their relations to topics. If you only want to see this page you should disable all else when calling visualize().

topicwizard.visualize(corpus=texts, pipeline=pipeline, exclude_pages=["topics", "documents"])

Word map

On the left you will see a plot showing you all the words, aka. the word map. On the map positions are determined by 2-dimensional UMAP projections of the transposed components of the topic model.

Selecting Words

You can select words by typing them into the field on the top left and searching for them. Multiple words may be selected at the same time. Clicking a word on the map adds the word to the selection.

Associations

Closely associated words also get highlighted on the graph and are included in calculations. The most closely associated words are the ones that have the lowest distance to the given words. Topicwizard finds the specified amount of closest words to the selected ones. The number of associated words can be adjusted with this slider:

Important Topics

You can see which topics use the selected word and their associations most frequently by glancing at the bar plot to the right.

Self-Contained Plots

It might be an overkill for you to display the entire page, and you might want static html plots instead of the entire application running. This can be particulary useful for reports with DataPane or Jupyter Notebooks.

Word Map

The word map that you can display with a dedicated function is slightly different from the one in the app as here you can’t select words to highlight.

Instead you can specify a cutoff in Z-values over which words will be labelled on the graph.

Words are also distinctively colored according to the most relevant topic as you cannot select the individual words for inspection.

from topicwizard.figures import word_map

word_map(corpus=texts, pipeline=pipeline)

Topics

You can visualize most relevant topics for a given set of words with barcharts, these behave virtually the same as in the app, but no associations are selected by default.

So for example if we would like to know which topics contain the words “supreme” and “court”, we can do so:

from topicwizard.figures import word_association_barchart

word_association_barchart(["supreme", "court"], corpus=texts, pipeline=pipeline)